• No results found

Essays on labour demand and wage formation

N/A
N/A
Protected

Academic year: 2021

Share "Essays on labour demand and wage formation"

Copied!
190
0
0

Loading.... (view fulltext now)

Full text

(1)

Publications

60

Essays on labour demand and wage

formation

(2)

60

Essays on labour demand and wage formation

Ossi Korkeamäki

Valtion taloudellinen tutkimuskeskus Government Institute for Economic Research

(3)

ISBN 978-952-274-011-3 (nid.) ISBN 978-952-274-012-0 (PDF) ISSN 0788-4990 (nid.)

ISSN 1795-3332 (PDF)

Valtion taloudellinen tutkimuskeskus

Government Institute for Economic Research Arkadiankatu 7, 00100 Helsinki, Finland Email: etunimi.sukunimi@vatt.fi

Edita Prima Oy

Helsinki, January 2012

(4)

Essays on labour demand and wage formation

Ossi Korkeamäki

Academic dissertation to be presented, by the permission of the Faculty of Social Sciences of the University of Helsinki, for public examination at Economicum, Lecture Room, Arkadiankatu 7, on February 3, 2012, at 12 noon.

(5)
(6)

Government Institute for Economic Research

VATT Publications 60/2012

Ossi Korkeamäki

Abstract

The first essay in this thesis is on gender wage differentials among manufacturing sector white-collar workers. The wage differential is decomposed into firm, job (within-firm) and individual-level components. Job-level gender segregation explains over half of the gap, while firm-level segregation is not important. After controlling for firm, job and individual characteristics, the remaining unexplained wage cap to the advantage of men is six per cent of men’s mean wage.

In the second essay, I study how the business cycle and gender affect the distribution of the earnings losses of displaced workers. The negative effect of displacement is large, persistent and strongest in the lowest earnings deciles. The effect is larger in a recession than in a recovery period, and in all periods women’s earnings drop more than men’s earnings.

The third essay shows that the transition from steady employment to disability pension depends on the stringency of medical screening and the degree of experience-rating of pension costs applied to the employer. The fact that firms have to bear part of the cost of employees’ disability pension costs lowers both the incidence of long sick leave periods and the probability that sick leave ends in a disability pension.

The fourth and fifth essays are studies on the employment, wage and profit effects of a regional payroll tax cut experiment conducted in northern and eastern Finland. The results show no statistically significant effect on any of the response variables.

Key words: Gender wage gap, gender segregation, displacement, earnings losses, disability pension, experience rating, payroll tax, tax incidence, labour demand

(7)

Ensimmäisessä esseessä miesten ja naisten välinen keskipalkkojen ero jaetaan sukupuolisegregaatiosta johtuvaan osaan ja henkilökohtaisista ominaisuuksista johtuvaan osaan. Tulosten mukaan yksityisen sektorin toimihenkilöillä yritys-tason segregaatio ei juuri ole osallisena palkkaeron syntyyn. Hiukan yli puolet kokonaiserosta selittyy naisten keskittymisellä matalapalkkaisiin töihin yritysten sisällä. Kun otetaan huomioon erot koulutuksessa, työkokemuksessa ja työn vaativuudessa on selittymätön palkkaero kuusi prosenttia miesten keskipalkasta. Toisessa esseessä näytetään, että irtisanomisen vaikutus ansiotulojakaumaan on pitkäkestoinen ja voimakkain jakauman alimmissa desiileissä. Irtisanomisen vaikutus tulojakaumaan on paljon suurempi laman aikana kuin kasvuperiodilla ja negatiivinen vaikutus naisten tuloihin on kaikissa oloissa voimakkaampi kuin miehillä.

Kolmannen esseen tulokset osoittavat, että työntekijän terveydentilan lisäksi myös lääketieteellisten kriteerien tiukkuus ja yritysten omavastuu eläkkeen kustannuksista vaikuttavat työkyvyttömyyseläkkeelle jäämisen todennäköi-syyteen. Vaikutus on kaksiosainen: omavastuu vähentää pitkiä sairaslomia ja kasvattaa työntekijöiden todennäköisyyttä palata takaisin töihin sairastumisen jälkeen

Neljännessä ja viidennessä esseessä tutkitaan Pohjois-Lapin ja Kainuun alueella toteutetun yritysten sotumaksuvapautuksen vaikutusta yritysten työllisyyteen, palkkasummaan ja voittoihin sekä työntekijöiden palkkoihin. Kokeilulla ei havaittu olevan tilastollisesti merkitsevää vaikutusta yhteenkään vaste-muuttujaan.

Asiasanat: Sukupuolten palkkaero, segregaatio, irtisanominen, työkyvyttömyys-eläke, yritysten omavastuu, palkan sivukulut, työn kysyntä, verotuksen kohtaanto

(8)

The foreword or acknowledgements section of a thesis usually describes the process of writing the thesis: where it all began, who and what were the main influences along the way – sometimes even what it all was like. In my case, were I compelled to condense everything to the minimum, I could do it in four letters: VATT. The Government Institute for Economic Research provided the necessary resources and supervision. It tolerated delays when I stalled and then pushed me forward again. It might seem impersonal to give a first round of thanks to an institute, but I believe that VATT has been more than the sum of the persons that work there. For this I wish to thank the Directors General Reino Hjerppe, Seija Ilmakunnas and Aki Kangasharju.

My first supervisor and superior at VATT was Research Director Pasi Holm. He was followed by Seija Ilmakunnas, Heikki Räisänen, Roope Uusitalo, and for the last leg by Anni Huhtala. I thank them for their encouragement and for their unwavering belief in my ability to finish this project. If there was any doubt, they kept it to themselves and kept me going... and going. I might well be the last economist to have started his thesis in the previous millennium and actually complete it. Roope is also co-author of one of the articles and became my academic supervisor towards the end of this venture. I am grateful for his help and advice. I would also like to thank Kari Hämäläinen. In his own words I was “unsupervisable”, but his door was always open and quite often I dropped in for advice anyway.

If one only scans the contents of my thesis it is obvious to whom the most gratitude is due. Tomi Kyyrä is co-author of three of the papers. He is a great colleague and without him I doubt this thesis would exist. Jukka Appelqvist worked with us on the displacement paper and Antti Luukkonen took part in the gender wage differentials project. Thank you Jukka and Antti. There are many other colleagues I want to thank for their good company and actual help: Juha Tuomala, Matti Sarvimäki, Teemu Lyytikäinen and Tuomas Kosonen. Whenever I needed help with computers, I could rely on Raimo Hintikka’s swift response. Sari Virtanen was essential in getting this document printed and Andrew Lightfoot helped me with the language.

The other vitally important institute for my research was Statistics Finland. I wish to thank the persons there working for the research laboratory: Mika Maliranta, Satu Nurmi, Jouko Verho, Antti Katainen and Marjo Pyy-Martikainen.

I would like to thank my pre-examiners, Professors Rudolf Winter-Ebmer and Oskar Nordström Skans, for their comments.

(9)

On a more personal note, I would like to thank Janice Redman, Tarja Nyberg and the Sparrows. Thank you Hannele, Joonas and Matias.

(10)

1.Introduction 1 1.1A gender wage gap decomposition for matched employer-employee

data 1

1.2A distributional analysis of earnings losses of displaced workers in an

economic depression and recovery 3

1.3Institutional rules, labour demand and retirement through disability

programme participation 5

1.4Employment and wage effects of payroll tax cut – evidence from a

regional experiment 7

1.5The Finnish payroll tax cut experiment revisited, or where did the

money go? 9

2.A Gender wage gap decomposition for matched employer-employee

data 13

2.1Introduction 13

2.2Methodological framework 16

2.2.1The wage model 16

2.2.2Decomposing the gender gap in pay 17

2.2.3The fixed effects approach 18

2.2.4The correlated random effects approach 19

2.2.5Discussion 21

2.3Data and descriptive statistics 23

2.3.1The TT data and job classification 23

2.3.2Sample statistics 25

2.4Results 30

2.4.1Wage regressions 30

2.4.2Wage gap decompositions 33

2.4.3Robustness of the results 36

2.4.4Comparisons with findings from other studies 40

2.5Conclusion 42

3.A distributional analysis of earnings losses of displaced workers in an

economic depression and recovery 45

3.1Introduction 45

(11)

3.3Data and sample construction 52

3.4Descriptive evidence 55

3.4.1Macroeconomic environment 55

3.4.2Background characteristics 57

3.4.3Empirical earnings distributions 59

3.5Quantile displacement effects 62

3.5.1Pre-displacement effects 71

3.5.2Post-displacement effects 73

3.5.3Robustness of the results 75

3.5.4Comparisons to results from mean regressions 77

3.6Concluding remarks 80

4.Institutional rules, labour demand and retirement through disability

programme participation 85

4.1Introduction 85

4.2Related literature 87

4.3Institutional framework of Finland 91

4.3.1Sickness and disability benefits 91

4.3.2Experience rating of disability pension benefits 93

4.4Data and descriptive evidence 95

4.4.1Incidence of disability retirement 96

4.4.2Outcome variables for analysis of transitions 97

4.4.3Sample design for modelling transition rates 100

4.4.4Raw transition rates 100

4.5Determinants of transition rates 104

4.5.1Individual characteristics 104

4.5.2Strictness of medical criteria 107

4.5.3Experience rating 108

4.5.4Employment growth and excess turnover 112

4.6Concluding remarks 115

5.Employment and wage effects of a payroll tax cut – evidence from a

regional experiment 119

5.1Introduction 119

(12)

5.5Data 129

5.6Results 131

5.6.1Covariate balancing 131

5.6.2Employment and wage sum responses to the regional payroll

tax experiment 134

5.6.3The effects by firm type 136

5.6.4The effect on wages 137

5.7Concluding comments 139

6.The Finnish payroll tax cut experiment revisited, or where did the

money go? 143

6.1Introduction and background 143

6.2The experiment, target and comparison regions and firms 145

6.2.1Finnish payroll taxes 146

6.2.2Target and comparison regions used in the evaluation 147

6.2.3Target and comparison firms 151

6.3Data sets 154

6.4Identification 155

6.5Results 157

6.5.1Where did the money go? 159

6.5.2Effect on employment, wage sum and profits 161

6.5.3Effect on wages and hours, individual wage records 163

6.5.4Robustness checks 166

(13)
(14)

1.

Introduction

This thesis contains five microeconometric studies where I explore wage formation and labour demand in the Finnish labour market. All the essays are based on linked employer-employee data.

The first paper is an investigation into gender wage differentials. The focus is on to what extent these differentials arise from the segregation of the labour market into men’s jobs and women’s jobs on the one hand, and from segregation of men and women into different firms, and hence industries, on the other. The second essay is a distributional analysis of the earnings losses of displaced workers. The gender aspect is also present in the second study as clear differences are found in how the earnings of men and women respond to losing one’s job. The second essay also shows that macroeconomic labour demand conditions have a huge impact on the individual’s labour market success after displacement. The third study concerns the effects of experience-rating firms’ disability pension costs on the incidence of disability pensions. There I show that the design of institutions involved with disability pensions has a strong effect on firms’ labour demand decisions. In the fourth paper, I evaluate the employment and wage effects of a regional payroll tax cut. The last study widens the scope of the evaluation to include the effects of the tax cut on the profitability of firms.

This introduction continues with a short description of each of the studies. I discuss the main results and conclusions and try to highlight the contributions to our knowledge of the Finnish labour market. The papers are independent in their surveys of previous literature, and the methods and datasets are also thoroughly described. Hence, rather than constructing the theoretical background, building the apparatus for analysis and describing the relevant institutions, I aim at brevity in this introductory chapter.

1.1 A gender wage gap decomposition for matched

employer-employee data1

In this paper the gender wage gap is decomposed based on a correlated random effects model. The decomposition makes it possible to assess the extent to which the overall gap is attributable to gender segregation at the firm level, gender segregation into jobs within firms and within-job wage differentials. The data set comes from the Confederation of Finnish Industries and covers large and medium-sized manufacturing sector firms.

1 Korkeamäki, O. – Kyyrä, T. (2006): A gender wage gap decomposition for matched employer-employee

(15)

By explicitly modelling the firm and job effects, the approach also proves to be informative regarding the sources of lower pay in predominantly female firms and jobs. The data contains a detailed measure of job complexity that makes it possible to compare jobs within firms. In this study I present the results for manufacturing sector white-collar workers. The working paper version also contains the results for blue-collar workers.

The difference in mean wages between men and women is approximately 22 per cent of men’s mean wage, to the advantage of men. Firm-level segregation accounts for only a small part, three percentage points, of the differential. At the firm level, no firm or firm-averaged worker characteristics are strongly associated with the gap. The difference in the share of women of a firm’s employees between men and women explains only half of a percentage point of the total difference in wages.

The majority of the gender wage differential, a little over half, is attributed to the disproportionate concentration of women in lower-paying jobs within firms. High-paid managerial jobs are mainly occupied by men, which explains one fifth of the total gap. Among other types of jobs, men are concentrated in positions with higher skill requirements – that explains a tenth of the gap. Apart from differences in skills and managerial ability, these parts of the gap may reflect discrimination through differential access to higher-paying jobs, or they may result from gender differences in preferences. Although the reasons for the preponderance of women in lower-paying jobs remain a puzzle, our findings highlight the importance of equal opportunities in education, hiring, and promotion.

However, the results also suggest that predominantly female jobs pay lower wages than predominantly male jobs even if they are associated with a similar level of average education, tenure and job complexity. In other words, jobs of equal worth are differently rewarded depending on whether they are occupied by men or women. Of course, one can always speculate how accurate the job complexity variable is, but if we assume that this measure is reasonably good, the results would imply that policies like comparable worth should be considered. One third of the wage differential arising from the level of jobs within firms is due to the “femaleness” of the job and cannot be explained by other variables. Finally, I found that within jobs women are paid some six per cent less than their equally qualified male co-workers. The unexplained within-job gap is higher among more educated and more experienced workers. Eliminating the sources of unexplained within-job wage differentials can directly account for a quarter of the overall gender gap in pay. On the one hand, this six per cent gap is a lot smaller than the often-quoted figure of 20 per cent drawn from gender averages; on the other hand, it is far from being insignificant. It is surprising that a gap of this size in hourly pay exists for equally qualified workers in the same job and

(16)

firm. The findings are in line with results from Sweden and Norway, but in Denmark the unexplained within-job wage gap is much larger – 14%.

The earliest gender pay differential studies concentrated mainly on personal characteristics, largely omitting the role of labour market segregation, whereas the first segregation-oriented papers used the female share at different levels of aggregation (industry, occupation, firm and job) as explanatory variables. In this study, the nested structure of the data is taken into account in the estimation and the results show that not doing so will give misleading results.

1.2 A distributional analysis of earnings losses of displaced

workers in an economic depression and recovery2

The gender wage differential study is, strictly speaking, a descriptive cross-section data story. The analysis is done meticulously, even obsessively so, but the dynamics leading to a certain wage distribution could only be guessed at. In the second study of this thesis, I look at a specific event in the labour market, displacement from a firm, and try to track its effects on the distribution of earnings over a period of time.

I analysed the earnings losses owing to involuntary job loss among Finnish workers who became displaced during a period of depression (1992) or recovery (1997). These groups of displaced workers and the associated comparison groups were followed over an 11-year period beginning three years before and ending seven years after the year of possible displacement. A few years ago, the early 1990s Finnish recession could have been considered an extreme case and not likely to have much relevance in more normal times. Now that the financial crisis has dealt a serious blow to a number of economies, the results could give some guidance to its possible labour market consequences.

Using the quantile regression method, I estimated the effect of displacement at each decile of the earnings distribution. The findings from both periods suggest that 1) displaced workers suffer from substantial and persistent earnings losses, 2) women are subject to larger earnings losses than men, and 3) the effect of displacement is very heterogeneous, being much larger in the lower quantiles and implying a sharp increase in the earnings dispersion following displacement. The fourth finding comes from the comparison of the effects of displacement at different phases of the business cycle: in a recession the earnings distribution of

2 Unpublished manuscript. Earlier working paper version: Korkeamäki, O. – Kyyrä, T. (2008): A

Distributional analysis of displacement costs in an economic depression and recovery. VATT Discussion Papers 465, Government Institute for Economic Research.

(17)

the displaced workers falls far below the counterfactual distribution; in a recovery period only the lower half of the earnings distribution is affected.

The first finding is in accordance with results from the US labour market. The results from other European labour markets are more mixed in this respect. The second result is interesting given that most US studies have not found notable differences between women and men, whereas the gender aspect is given very little attention in European studies. One possible explanation for the larger earnings losses of women could be that in Finland, as in other countries, women are more frequently out of work for family reasons. That may induce employers to favour male employees when investing in managerial and professional skills. If so, and if such skills are generally transferable, i.e. not lost in job displacement, one can expect to find smaller earnings losses for displaced men. Taken together with the fact that managerial, professional and technical jobs are disproportionately held by men, employers’ investment behaviour may lead to larger earnings losses for women. It remains unclear why earnings losses differ between men and women in Finland, but not in the US. It should be stressed that women’s labour market position is quite different in Finland. On the one hand, the relatively generous maternity and parental leave schemes encourage career breaks, but on the other hand, public day care and school meals help the mothers of young children to work full-time if they want to. Moreover, part-time work among women in Finland is not very common and the labour force participation rate of women is relatively high.

The third finding, the heterogeneity in the displacement effect, is the main contribution of the paper and has important implications. First, the positive effect on earnings dispersion means that job loss does not only cause a significant decline in expected earnings, but also creates uncertainty about the level of future earnings. This suggests an additional welfare loss for risk-averse workers. This effect has typically been ignored in the discussion of displacement costs. Secondly, the large effect at the lower end of the distribution is consistent with the hypothesis that the relative importance of transferable individual-specific skills, which are not lost in job displacement, is larger for high-ability workers, who tend to populate the upper part of the conditional earnings distribution. Finally, the disproportionately large effect on the first two deciles implies that the effect on the expected earnings loss is in large part driven by an increased risk of joblessness and low-paid employment following job displacement. This means that job training and job replacement programmes targeted at unemployed job seekers, if effective in enhancing re-employment, can provide a means to reduce the average displacement cost.

By comparing the results from the two periods (displacement in 1992 or 1997), we found much larger earnings losses for those who lost their jobs during the depression period. Men (women) who were displaced in the middle of the depression had approximately 58% (65%) lower median earnings one year after

(18)

job loss and 15% (20%) lower median earnings seven years after job loss. Because of the exceptionally difficult labour market conditions, their earnings distribution as a whole remained below the counterfactual level through the end of the follow-up period. By contrast, job loss in the recovery period had a long-lasting effect only on the lower half of the distribution. For workers displaced in 1997 the effect on the median of the earnings distribution of men was 3.9% (women 7.1%) one year after being displaced. Seven years later the effect was still approximately 8% for women, whereas it did not differ from zero for men. These long-term losses do not vanish even if we account for income transfers.

1.3 Institutional rules, labour demand and retirement through

disability programme participation3

The disability benefit scheme is one of the largest social security programmes in many countries, and therefore is of particular interest. In Finland, disability is the most common reason for early retirement, and disability expenditure accounted for some 3.5% of GDP in 2003, which was the third highest share in the EU after Sweden and Denmark. Disability enrolment rates of older employees vary strikingly across the European countries and the US. These cross-country differences cannot be explained by demographic or health-related factors. Over the past two or three decades, many countries have also experienced an expansion of disability benefit enrolment even though their ageing populations have become healthier. This is a serious concern given the common goal of inducing people to retire later. The widespread use of disability benefits as an early retirement instrument has been argued to be a particularly serious problem in Finland.

When job cuts are necessary, firms often offload their oldest employees first. If the health requirements for disability benefit eligibility are weak, early retirement via the disability scheme can be a useful strategy in effective downsizing, providing a way to reduce the workforce in a “soft” way. Some firms may also target dismissals at those employees with a high risk of disability. In doing so, the employer can avoid disability costs arising from the experience-rated contributions of disability pension benefits. Encouraging disability retirement could also be an attractive strategy for an employer wanting to change the composition of the workforce at a time of stable or growing employment when dismissals are difficult to justify.

Much previous empirical literature has been based on a simple labour supply framework in which an employee chooses whether to apply for disability

3 Korkeamäki, O. – Kyyrä, T. (2011): Institutional rules, labour demand and retirement through disability

(19)

benefits, while the employer has no role at all. Surprisingly little effort has been made to study the labour demand side. This essay aims to shed light on the relationships between labour demand, institutional factors and early retirement through disability programmes. I consider the importance of the labour demand side by examining the relationship between the growth and restructuring rates of an establishment and entries into disability. In addition, I assess the effectiveness of two policy instruments: the strictness of medical requirements for disability pension eligibility and the experience-rating of disability expenditure. The first determines the ease of access to disability pension benefits, whereas the latter places part of the costs of early retirement on the employer.

Transitions out of work to sick leave and disability retirement are modelled using matched employer-employee data for the Finnish private sector covering the years 1991–2005. The data set includes all active firms and employees can be tracked across all labour market states. To identify the role of institutional factors I exploit a law change that made the medical requirements for disability pension eligibility tougher for a certain group, as well as changes in partially experience-rated employer contributions.

The main findings can be summarized as follows. 1) For older employees a transition to sick leave is often a one-way street out of employment, leading eventually to disability retirement. Half of 50−55 year-olds and over two-thirds of older workers on sickness benefits end up in disability retirement within the next three years. This highlights the importance of preventive measures aimed at minimizing the flow into sick leave. 2) Those employees who could apply for a disability pension under more lenient medical requirements were much more likely to enter sick leave and to retire via disability pension benefits. Therefore, the abolition of the individual early retirement scheme in 2000 significantly reduced the flow into disability retirement in the affected groups. 3) I find strong evidence that experience-rating lowers the flow into sick leave and reduces transitions from sick leave to disability retirement. Moreover, those large firms that can easily bear their share of early retirement costs owing to their strong financial position more readily let employees who are already on sickness benefits exit via disability pension schemes than firms in a weaker financial position. Financial strength does not matter for smaller firms that are not subject to experience-rating. 4) The transition rates to sick leave and disability retirement are relatively large in establishments experiencing a high degree of excess worker turnover. When an establishment is growing, transitions to sick leave and disability retirement become less frequent. There is no evidence that employers exploited the disability pension scheme as a way of adjusting their workforce when downsizing.

These findings imply two policy recommendations to reduce the disability benefit enrolment rate of older workers. First, the stringency of medical criteria and medical screening for disability benefit eligibility should be tough enough.

(20)

When non-medical factors are weighted at the expense of medical criteria, disability benefits may distort labour supply decisions, thereby also inducing workers who are not truly disabled to retire via disability programmes. This appears to be mainly a labour supply issue, as I did not find evidence that employers encouraged disability retirement when downsizing. Secondly, the experience-rating of disability benefit costs seems to be an effective policy instrument. It probably induces employers to take preventive action to reduce the inflow into sick leave. Firms also put more effort into getting employees on sickness benefits back into work. This finding should be of considerable interest, not only for Finland, but also for the other countries that do not yet have an experience-rating system for disability benefits. Obviously, there are still a number of open questions regarding, for example, the optimal design of experience-rating and possible spillover effects on hiring and transitions out of work to other destinations than disability retirement.

1.4 Employment and wage effects of payroll tax cut – evidence

from a regional experiment4

In this paper, I evaluate the employment and wage effects of a regional tax cut experiment in northern Finland. The experiment started in 2003 and it was due to continue for three years, but it was soon extended to 2009 and then again to the end of 2012. The experiment abolished employer contributions to the national pension insurance and national health insurance schemes for firms located in certain high-unemployment regions. Prior to 2003, these employer contributions varied between 2.95 and 6 per cent of the wage bill, depending on the capital intensity and size of the firm. The average payroll tax reduction of the experiment was 4.1 percentage points.

The evaluation setup was designed well before the start of the experiment. First, I chose a comparison region closely resembling the target region and then continued by matching the target region firms to the comparison region firms to form groups of firms as comparable as possible. I did this step at an early stage to make the evaluation transparent and credible. The timing was not very fortuitous, however. The Kainuu self-government experiment, featuring a similar reduction in payroll taxes, started only two years after the experiment in Lapland – and its target area was in the middle of the comparison region. Hence this study considers only the first two years of the experiment.

The main result was that the tax cut did not have a statistically significant effect on employment or the firms’ wage sum. The wage sum seemed to have risen but

4 Korkeamäki, O. – Uusitalo, R. (2009): Employment and wage effects of a payroll tax cut – evidence

(21)

the standard errors of the estimates were too large to warrant any strong conclusions. These results were based on firm data acquired directly from the tax authorities. I also had data on individual wages from the employer associations. If the estimates from the sub-sample of firms for which wage data is available can be generalized to all firms, about half of the effect of the payroll tax reduction on labour costs was offset by faster wage growth. The remaining two per cent decrease in labour costs did not have a significant effect on employment. According to my estimates, the demand and supply elasticities are roughly equal. The point estimate that the tax cut increased employment by 1.3 percent indicates labour demand elasticity of around 0.6. This is well within the range of earlier estimates. Unfortunately the confidence bands around this estimate are too wide to give much guidance for future tax policy.

The results are in line with findings from other micro-level empirical studies. Usually no employment effects are found and if any effect exists, it is a partial shifting of the tax cut to wages. The most relevant comparisons are with Sweden, where two payroll tax cut experiments were conducted. The first started in 1982 in the four northernmost municipalities and was eventually extended to cover the whole of regional support area A, i.e. almost the northern half of the country, excluding the coast. Payroll taxes were cut by ten percentage points and there was no ceiling to the cut. The experiment, which had become a semi-permanent regional subsidy, had to be phased out by the end of 1999, owing to EU regulations. The second experiment, started in 2002, was similar in geographical scope. It also reduced taxes by ten percentage points but the maximum deductible amount was rather low. Neither of the experiments yielded a statistically significant employment effect. The effect on wages was not investigated for the first experiment and it was positive and significant in the latter case.

The fact that the cut in payroll taxes was targeted at narrowly defined regions and the temporary nature of the tax cut limit the extent to which the results can be generalized to the potential effects of a permanent country-wide reduction in payroll taxes. First, the payroll tax experiment was financed by increasing payroll taxes in the rest of the country. In a national scheme, the budgetary cost would need to be financed by raising other taxes. Second, a regional experiment may have substitution effects if firms reallocate labour to the target region from the rest of the country. This might be beneficial in the sense that part of the reasons for the regional payroll tax cut was to boost employment in disadvantaged regions. However, this limits the usefulness of the results from the experiment in predicting the effects of a national programme. Third, the incidence of the tax cut may also be different in a regional programme since wage contracts are negotiated at the national level. Any nationwide changes in payroll taxes may have an impact on the outcome of these negotiations, while a regional programme that only affects a small share of employers has little weight in national bargaining. Finally, a temporary programme is likely to create smaller employment effects than a permanent reduction in payroll taxes. The expected

(22)

duration of three years might not have been a sufficiently long period for firms to adjust their labour demand to a relatively small change in labour costs.

1.5 The Finnish payroll tax cut experiment revisited, or where

did the money go?5

The payroll tax exemption was originally planned to last for three years, from January 2003 to December 2005. Already in May 2003, the government had decided to start a regional self-government experiment in Kainuu, eastern Finland, beginning from 2005. That experiment contained a similar provision for lowered payroll taxes as the Lapland experiment. Hence it spatially enlarged and temporally extended the payroll tax experiment until the end of 2009. The experiment has been further continued until the end of 2012 and there is intense lobbying to make it a permanent arrangement.

The first evaluation study was somewhat stunted by the start of the Kainuu experiment. The results from the first two years were in the expected direction but less than satisfying in their precision. For this study I had both more firms in the treatment group and more years of observations. I also had information on firms’ balance sheets and financial statements that allowed me to track whether the tax cut had an effect on firms’ profits, if not on employment or wages. To sum up the situation at the start of writing the fifth essay and to explain the need to re-evaluate the effects of the payroll tax cut. I stated the following:

1) According to our previous research on the first two years of the payroll tax experiment, the tax cut in northern Finland does not seem to have had an immediate employment effect. This finding is consistent with evidence from other Nordic labour markets. 2) In the earlier study there was some indication of rising wages, but not 1:1 with respect to the tax break – this is also a common finding from Sweden and Norway. From 1) and 2) and supported by results from the UK, where a minimum wage change had a negative effect on profits (but no employment effects), it seems likely that changes in payroll tax could also have an effect on firm profitability. Models of incomplete competition from the industrial organisation literature and matching models from the labour market side can accommodate these profit effects, but their size remains an empirical question.

With the larger target area for the experiment, the evaluation setup of the previous study was no longer valid – the Kainuu region formed the main part of

the comparison region. The target regions in northern Lapland and Kainuu are

not geographically linked, but both are within the region eligible for the highest

5 Korkeamäki, O. (2011): The Finnish payroll tax cut experiment revisited, or where did the money go?

(23)

national firm subsidies. From 2000 to 2007, Kainuu was in the highest subsidy region and northern Lapland was part of the region eligible for the second-highest subsidies. There was, however, a special provision that granted firms in Lapland access to subsidies almost as high as for firms in the first category. Instead of doing another matching exercise, I used firms within the same subsidy region but outside the target region of the payroll tax cut as a comparison group (see Figure 6.1).

The target group consists of approximately 2900 firms that in 2001 employed an average of 3.4 workers and had an average turnover of 466,000 euros. The firms in the comparison area were slightly larger, having 3.7 employees and a turnover of 497,000 euros. Prior to the experiment the comparison area firms had also grown somewhat faster than the experiment region firms, but none of the firm-level pre-experiment response measures (firm-levels or growth) differed statistically significantly between the firm groups, even before controlling for industry, growth trends, etc. Therefore I claim that the setup was rather successful. The analysis was done with differences-in-differences regressions, with controls for either industry- and region-specific or firm-specific growth trends.

The main result from this new study is that the payroll tax cut did not have a statistically significant effect on employment, the wage sum, profits or hourly wages in the private sector. Most of the estimates are positive but unfortunately the standard errors are so wide that they could accommodate values indicating a full shifting of the tax cut to either the wage sum, profits or, indeed, to employment. If we look at the point estimates in euro or employee terms, they indicate that the wage sum in the target region firms rose faster than in the comparison region, employment growth did not react and profits grew even a little more than the wage sum. Alternatively, if we consider the point estimates of the percentage changes, employment and the wage sum did grow by an equal amount and there was no effect on profits. The only unambiguous finding is that the tax cut is reflected in the financial statement data, although, even there, there was some uncertainty in the case of small and the least capital-intensive firms. The effect on hourly wages found in our earlier study is not found for the combined target region of Kainuu and Lapland. The effect is still found for Lapland – but the estimates for Kainuu would imply a negative wage effect and the effects cancel each other out when calculating the total. The results also show one statistically significant change in a non-experiment year that calls for caution in interpreting the results. Certainly, a region-specific shock could have taken place in Kainuu and caused the negative effect, but I found no reason to believe that the result for Lapland is trustworthy.

Irrespective of the findings of this and other similar experiments in other Nordic countries, national pension insurance payment contributions have gradually been lowered in recent years on the premise that this is a cost-effective way of

(24)

boosting employment. From the beginning of 2010, they were abolished altogether. There was some debate as to whether this was the most effective way to help firms to generate jobs, but empirical findings had a rather minor role in the discussion. This is, of course, partly due to the lack of conclusive findings. The Finnish payroll tax experiment is a rare example of a tax change being made in an experimental setting with the stated aim of facilitating economic research. Hence it was important to evaluate the experiment, even if the results tell rather little about the effects. This was also an opportunity to learn about the experiment itself in order to understand better how possible future experiments should be designed and implemented to the greatest scientific advantage. I argue that it is still important to continue experimenting – it is also important to pre-evaluate experiments to see if they are likely to yield accurate and reliable results.

(25)
(26)

2.

A Gender wage gap decomposition for matched

employer-employee data

6

Abstract

In this paper, we evaluate the extent to which the gender wage gap in the Finnish manufacturing sector is attributable to within-job wage differentials, gender differences in individual qualifications, and a disproportionate concentration of women in lower-paying firms and lower-paying jobs within firms. We use matched employer-employee data to compare wage differentials between similarly qualified female and male workers who are doing similar work for the same employer. Our modelling approach employs a correlated random effects specification to account for the hierarchical grouped structure of the underlying data.

Key words: Gender wage gap, wage discrimination, gender segregation, random effects model

JEL classification numbers: J14, J23, J26

2.1 Introduction

A huge body of literature has emerged to explain why the gender wage gap persistently exists in virtually all labour markets (see Altonji and Blank, 1999, and Blau and Kahn, 2000, for recent surveys). Traditional attempts to explain the wage gap focused on gender differences in individual qualifications and their rewards in the labour market. More recently, the importance of the segregation of women and men into different jobs has been recognized. This line of research emphasizes that wages are closely tied to the characteristics of jobs, not only to the individuals who hold them. If typical female jobs pay lower wages than jobs dominated by men, the mean earnings of women can fall short of men’s earnings even in the absence of within-job wage differentials between the sexes.

6 Ossi Korkeamäki and Tomi Kyyrä.

The earlier version of the paper was circulated under the title “Explaining gender wage differentials: Findings from a random effects model”. We appreciate the helpful comments received at the third Nordic Workshop on the Economic Analysis of Linked Employer-Employee Data in Bergen, the second Nordic Econometrics Meeting in Bergen, the CAED Conference in London, and the EALE Conference in Seville. We are grateful to the Confederation of Finnish Industry and Employers for access to their data. Antti Luukkonen kindly provided his measures of job complexity levels for our use. The suggestions of a co-editor and anonymous referees considerably improved the paper.

(27)

Attempts to quantify the segregation effects on the wage gap were for a long time distorted by a lack of appropriate data. Consequently, most early analyses focused on segregation among occupations, firms, or industries only. This is clearly unsatisfactory, as women and men are further segregated into different jobs within firms. In recent years, important advances have been made by access to large matched employer-employee data sets that contain multiple observations on workers with the same employer. When information on occupations or job titles is available, such data enable wage comparisons between male and female workers who are doing similar work for the same employer. This kind of comparative analysis has been conducted by Petersen and Morgan (1995), Petersen et al. (1997), Meyersson Milgrom et al. (2001), Groshen (1991), Datta Gupta and Rothstein (2001), and Bayard et al. (2003). In the first three of these studies observed sex differentials in mean wages within jobs are simply aggregated to form various wage decompositions. This approach has the obvious drawback that variation in individual characteristics is left uncontrolled. In the other studies, wages are regressed against a set of control variables and fraction female in the worker’s industry, firm, occupation, and/or job.7 The key idea is that the regression coefficients of the various fraction female variables capture the relationship between the wage rate and ‘femaleness’ of the underlying labour market structure.

It should be noted that a common practice in the fraction female regressions above has been to neglect the grouping in the underlying data. For example, observations on workers resulting from the same firm are interpreted as being independent.8 However, intuition suggests that we should expect workers in the same firm to be more homogeneous than those in a sample drawn randomly from the population of all firms. Workers in the same firm share many common factors, some of which may be observable (e.g. firm size, fraction female) but many are not (e.g. market power, managerial ability). In the regression analysis the effect of such unobservables serves as a latent firm effect that will be absorbed into the error term. Moreover, since different jobs require different skills and qualifications, we can further expect that within a given firm workers who are doing the same job are more homogenous than the firm’s workforce as a whole. This implies an additional source of dependence between workers within jobs.

In general, the matched employer-employee data exhibit a particular type of grouped structure, which contrasts the statistical properties of such data with the classical random sample case. A consequence of the grouping in the regression

7 In a related paper, we apply this method to the Finnish data; see Korkeamäki and Kyyrä (2002).

Groshen’s (1991) specifications do not include control variables.

(28)

analysis is that, owing to the latent group effects, the errors will be correlated within groups. In the absence of correlation between the latent group effects and regressors included in the model, the OLS coefficients will be consistent, but the usual standard error estimates can be very misleading (Moulton, 1986). More generally, when the group effects are correlated with the regressors, the OLS coefficients will be inconsistent.

In this study, we explore wage differentials between white-collar women and men in the Finnish manufacturing sector using a large matched employer-employee data set. We view the data as having a nested structure with three levels: firms, jobs within firms, and workers in jobs within firms.9 A job is defined as an occupation within a firm. Along with individual characteristics, the wage rate is allowed to depend upon the employing firm and the job the worker holds within the firm. The latent firm and job effects are modelled as functions of group characteristics, including the mean characteristics of individuals within the groups. We end up with a regression model with variables measured at the individual, job, and firm levels, and an error term that has a two-way nested structure with separate intercepts for firms and jobs within firms. Using the regression results we decompose the overall gender gap in pay into the contributions of gender segregation, gender differences in the individual qualifications, and the unexplained within-job gap.

Our approach departs from the existing segregation literature in that we explicitly model wage differentials between firms and jobs. In contrast to standard fraction female regressions, we obtain consistent estimates of the parameters of interest in the presence of the correlated group effects that are likely to arise in the case of the matched employer-employee data. With respect to job segregation, the previous studies have focused on quantifying what fraction of the overall wage gap can be attributed to a disproportionate concentration of women in lower-paying jobs. In addition to identifying this quantity under less restrictive assumptions, we go a step further by addressing the issue of why typical female jobs are lower-paid. When evaluating the extent to which lower wages in predominantly female jobs can be explained by job attributes, we make use of an index of job complexity that measures the responsibility, skills and effort required by a given job. Thus we are able to assess whether wage differentials between typical female and male jobs can be viewed as justified or not, a question that is beyond the scope of earlier analysis but crucial, for example, in the view of comparable worth policy.

9 Obviously, we could go further and introduce an additional level on top of this hierarchy by grouping

firms by industry. For simplicity, we focus on the three levels and treat industry as a characteristic of firms rather than a hierarchy level of its own.

(29)

In the next section, we describe the econometric methods and contrast our approach with the fraction female decompositions in the previous studies. Section 2.3 gives details about the data and reports some descriptive statistics. The results are reported in Section 2.4, which is followed by a concluding section.

2.2 Methodological framework

2.2.1 The wage model

Suppose our data consist of all employees of F firms. Within firms employees who do similar work are grouped together, in which case they are said to hold the same job. Observations across firms are regarded as being independent, but within firms wages are correlated owing to common firm and job characteristics. We model the log wage of worker i (i = 1, 2,…, njk) who holds job k (k = 1, 2,…, cj) in firm j (j = 1, 2,…, F) as

,

jki jki jki j jk jki

w =ηs +βx + +f v +ε (1)

where s is the female dummy, x is a vector of other individual characteristics, ν is the job effect that is nested within the firm effect f. For the idiosyncratic errors ε, we assume

( jki j, j, j 0 and ( jki jki j, j, j 0 for ,

E ε X v f = E ε ε X v f = i i≠ ′ (2)

where Xj includes x and s for all employees of firm j, and vj =(νj1, νj2,…,νjcj). Wage variation between firms and jobs beyond the observable individual characteristics is captured by f and ν respectively. Without loss of generality, the job effects are defined in deviation from the firm effects, with the expected value within each firm equal to zero. Thus, E(f + ν | firm j) = fj + E(ν | firm j) = fj. We emphasize that f and ν are likely to be correlated with s and x. In particular, women are expected to be concentrated in firms with low values of f, and further in jobs with low values of ν. Since different firms and jobs require different qualifications, the group effects f and ν are likely to be correlated also with the variables in x. If fj > fj’, workers in firm j earn more on average than workers in firm j’ after controlling for s and x. Similarly, provided that νjk > νjk’, workers in job k are more highly paid on average than those in job k’ within the same firm j, after controlling for s and x.

Within jobs wage differentials are related to workers’ sex (s), other individual characteristics (x), and unobservables (ε). A parameter of particular interest is η

that gives the expected wage differential between equally qualified (in terms of x) women and men who are doing the same work for the same employer. One

(30)

may be tempted to view a negative value of η as evidence of wage discrimination against women. Such an interpretation is justified only if all relevant explanatory variables were included in x. This may not be the case in practice. In general, the influences of possible discrimination and unmeasured individual characteristics are indistinguishable in the value of η. Therefore, we interpret η simply as a measure of the unexplained within-job wage differential between sexes.

At this point, a few remarks on the restrictions imposed above are in order. First, the returns to individual qualifications, β, are assumed equal for women and men. One should recognise that the interpretation of β is conditional on the position held in the labour market (i.e. conditional on f and ν), so β measures the returns within a given job. Since employers cannot apply very different reward schemes to their female and male employees who are doing the same work, our assumption is not as restrictive as it might first look. We will return to this issue and present results from a regression model with gender-specific slopes. The assumption that the unexplained within-job wage gap is of the same size everywhere is rather restrictive. One might wish to allow the coefficient of the female dummy to vary across jobs, i.e. replace η with ηjk. We adopt a very narrow definition for jobs in our empirical application. This results in a huge number of jobs, many of which include either female or male employees only, making the estimation of job-specific coefficients infeasible in practice.

2.2.2 Decomposing the gender gap in pay

The gender wage gap is defined as the difference in the expected wages between men and women, i.e. the wage difference between a randomly chosen man and woman. Using the model outlined above we decompose it as

[

]

[

]

[

]

( | 0) ( | 1) ( | 0) ( | 1) ( | 0) ( | 1) ( | 0) ( | 1) , E w s E w s E s E s E f s E f s E s E s η ν ν ′ = − = = − + = − = + = − = + = − = x x β (3)

where the contributions of gender segregation among firms and jobs are captured by the last two terms. A positive value of E(f | s = 0) – E(f | s = 1) indicates that women are disproportionately concentrated in lower-paying firms. This term would be zero, if there were no variation in f across firms or if women and men were identically distributed across firms. If women are relatively more frequently allocated to lower-paying jobs within firms, E(ν | s = 0) – E(ν | s = 1) will take a positive value. It would be zero, if there was no systematic wage variation across jobs within firms beyond the differences in individual characteristics or if, within all firms, women and men were allocated identically across jobs. The amount of within-job wage differentials between sexes not accounted for by the explanatory variables x equals –η. The contribution of gender differences in individual characteristics is captured by the remaining term on the right-hand side.

(31)

To obtain an empirical counterpart of the decomposition, the conditional means of w and x can be replaced with the sample means over women and men but the other components need to be estimated. Since the latent group effects are expected to be correlated with the explanatory variables, we will focus on estimation by fixed effects and correlated random effects.

2.2.3 The fixed effects approach

In our first approach, we take f and ν as fixed constants to be estimated along with η and β. Therefore, we consider the model conditional on the firm and job effects:

( jki| j, , )j j jki jki j jk.

E w X v f =ηs +βx + +f v (4)

In this case, η and β could be estimated by regressing w on s, x and the full set of job dummies. As the number of job dummies may be too large to make estimation feasible, we obtain analytically equivalent estimators of η and β by applying pooled OLS to the transformed model:

(

) (

)

,

jki jk jki jk jki jk jki jk

w w η s sε ε

⋅ ⋅ ⋅ ⋅

− = − +β xx + − (5)

where wjk, sjk, xjk, and εjk denote averages over workers in the k-th job of firm j. Under the assumptions (2), the resulting “fixed effects” (FE) estimators ηˆ

and ˆβ are consistent under arbitrary correlation between (s, x) and (f, ν). Given the restriction E(ν | firm j) = 0 for all j, the firm and job effects can be estimated as ˆ ˆ ˆ , j j j j f w ηs ′ ⋅⋅ ⋅⋅ ⋅⋅ = − −βx (6) ˆ ˆ ˆ ˆjk jk jk jk j, v =w ηs βx f (7)

where wj⋅⋅, sj⋅⋅and xj⋅⋅ denote averages over the employees of firm j. The point estimates of f and ν are noisy because the number of observations per firm and, especially, per job can be small. However, the estimates of their expected values among women and men based upon sample averages are expected to be reasonably accurate. Thus, we proceed by inserting ηˆ and ˆβ along with the sample means of ˆf and ˆv over women and men into (3). This gives the first version of our wage gap decomposition. It allows us to distinguish the contributions of gender segregation among firms and jobs to the overall wage gap from the contributions of the unexplained within-job gap and gender differences in individual characteristics.

(32)

2.2.4 The correlated random effects approach

In an alternative approach, we take an explicit account of the relationship between the latent group effects and the explanatory variables. More precisely, we specify the expected values of f and ν conditional on observables via auxiliary linear regressions. Let Xj* = (Xj, zj, gj1,…, gjcj) be the extended set of conditioning variables that includes firm attributes zj (firm size, industry, etc.) and job attributes gjk’s (job size, job complexity index, etc.) in addition to Xj. We specify the conditional mean of the firm effect as

0 1 2

( j | j) j j j

E f X= +α δ s⋅⋅ +δx ⋅⋅+δz (8)

and that of the job effect as

(

) (

) (

)

0 1 2 ( jk | j) jk j jk j jk j , E vθ s s ′ ′ ⋅ ⋅⋅ ⋅ ⋅⋅ ⋅ = − + − + − X θ x x θ g g (9)

i.e. the first moments of the marginal distributions of f and ν are assumed to be linear functions of the group means of s and x and of other group level variables. All the explanatory variables on the right-hand side of (9) are measured in deviation from the firm mean in order to enforce the expected value of ν within firms to zero.10

Now we consider the model conditional on Xj*:

( jki| j) jki jki ( |j j) ( jk | j).

E w X=ηs +βx +E f X+E v X

Defining ξjfj – E(f | Xj*) and ωjk = νjk – E(νjk | Xj*), we obtain the estimating wage equation:

0 0 1 1

2 2

( ) ( )

( ) ,

jki jki j jk j jki j jk j

j jk j jki w s s s s u α η δ θ ′ ′ ′ ⋅⋅ ⋅ ⋅⋅ ⋅⋅ ⋅ ⋅⋅ ′ ′ ⋅ = + + + − + + + − + + − + x x x x z g g β δ θ δ θ (10)

where ujki ξj + ωjk + εjki. Conditional on Xj*, all components of ujki are assumed to be mutually independent, with zero means and constant variances σξ2, σω2, and

σε2 respectively. Within firm j, the variance-covariance structure of the errors is given by

10 This is only a matter of parameterization provided that

j

(33)

2 2 2 2 2 2 , if and ; ( | ) , if and ; , if and . jki jk i j k k i i E u u k k i i k k i i ξ ω ε ξ ω ξ σ σ σ σ σ σ ∗ ′ ′  + + = =  ′ ′ = + = ≠  X (11)

This is known as the two-way nested error structure in econometrics (Fuller and Battese, 1973). It models the residual correlation within firms that remains after conditioning on the observed firm, job, and individual characteristics. Such a correlation is likely to exist owing to unobservable job and firm factors. We estimate the model with generalized least squares (GLS) that exploits the particular form of the error structure for efficiency and produces appropriate standard errors.11 It should be stressed that including the group means of individual explanatory variables in (8) and (9) provides a way of allowing s and x to be correlated with f and ν, an old idea by Mundlak (1978).12 To emphasize this point, we refer to the specification outlined above as the “correlated random effects” (CRE) model.13

Coefficients of the fraction female variables in (8) and (9) are of particular interest. A negative value of δ0 implies that firms with a high density of female workers pay lower wages after controlling for xj⋅⋅ and zj. If within firms employees in predominantly female jobs are lower paid given (xjkxj⋅⋅) and

(gjkgj), it will be indicated by a negative value of θ0. In other words, δ0 and θ0 are kind of “residual gender effects”, which imply that predominantly female firms and jobs pay different wages for reasons not accounted for by the observed worker and group characteristics.

Because E(f | s = 0) = E[E(f | X*) | s = 0] by the law of iterative expectations, we obtain an estimate of E(f | s = 0) by averaging the right-hand side of (8) over all men. E(f | s = 1) is estimated analogously by averaging over women. The contribution of gender segregation among firms can then be expressed as

11 The large unbalanced data raise some computational issues, as the inverse of the error

variance-covariance matrix is required by the GLS procedure. These issues and the estimation of the variance components are discussed in Korkeamäki and Kyyrä (2003).

12 Chamberlain (1984) considers a general case where the latent group effects are modelled as linear

predictors of s and x of all employees within the group. Mundlak’s (1978) specification is obtained by

imposing a restriction that the coefficients of s and x in the linear predictor are identical for all i within

the group. The unrestricted specification becomes cumbersome in our case where group sizes vary and some firms are very large.

13 The model defined by (10) and (12) is known under a variety of other names, including the nested error

components model, variance components model, random intercepts model, mixed model, and hierarchical model.

(34)

0 1 1 2 1 ( | 0) ( | 1) ( ) ( )( ), F m f j j j j F m f j j j j j E f s E f s o o s o o δ ⋅⋅ = ′ ′ ⋅⋅ = = − = = − + − +

δ x δ z (12)

where ojf (ojm) is the fraction of all women (men) allocated to firm j. Similarly, the contribution of gender segregation among jobs within firms is given by

0 1 1 1 2 1 1 ( | 0) ( | 1) ( ) ( ) ( ) ( ) ( ) , j j c F m f jk j jk jk j k c F m f jk jk jk j jk j j k E v s E v s o o s s o o θ ⋅ ⋅⋅ = = ′ ′ ⋅ ⋅⋅ ⋅ = = = − = = − −   + − − + −





θ x x θ g g (13)

where ofjk (omjk) is the fraction of all women (men) allocated to the k-th job of firm i. Substituting (12) and (13) into (3) along with the GLS estimates of the regression coefficients gives us the second version of our wage gap decomposition.

In the case of the CRE model, the segregation contributions can be expressed as sums of various terms. These terms pass on useful information, which is not available from the FE model. For example, if typical female jobs are found to be characterised by low values of ν, one may speculate that lower wages in such jobs result from lower skill requirements. If this is the case, a large fraction of the contribution of gender segregation among jobs in (13) will be attributed to differences in the mean education (incorporated in xjkxj⋅⋅) and job complexity (incorporated in gjkgj ), while the component associated with the fraction female (sjk⋅−sj⋅⋅) will be close to zero. By contrast, if wage differentials between typical female and male jobs arise largely from some unobserved sources, this will be indicated by a strong effect of the fraction female term in (13).

2.2.5 Discussion

In the case of the FE model, we cannot say anything about why predominantly female firms and jobs are lower paid on average. This of course is a cost of the robustness of the fixed effect method: we do not assume anything about the relationship between the group effects and regressors. Compared with the FE model, the CRE specification is more restrictive, as the conditional expectations of f and ν are assumed linear. However, when the group means of s and x are included in (8) and (9), the GLS estimators of η and β are identical to their FE estimators, and hence not affected by these additional restrictions. In this respect, we do not lose anything by imposing more structure on the model. The additional

(35)

structure of the CRE model is exploited in explaining wage variation between firms and jobs. While both the FE and CRE model are able to produce identical results for the effects of individual-level regressors, only the latter is informative about wage differentials between firms and jobs. For this reason, the CRE model is our preferred choice. A potential drawback of the method is that the sum of various contributions does not necessarily equal the raw wage gap in finite samples. This is a consequence of the more complex error structure.

Our approach departs from the decomposition exercises in Groshen (1991), Datta Gupta and Rothstein (2001), and Bayard et al. (2003) in some essential ways. Of course, the main difference is that our CRE approach is informative about the determinants of lower wages in predominantly female firms and jobs. Secondly, the interpretation of the regressor coefficients η and β comes from the wage model defined in (1) and (2), i.e. they measure wage differentials within jobs (In other words, the firm and job effects held constant). This interpretation is trivial when the model is estimated by fixed effects, but the coefficients have the same meaning also in the CRE specification as we allow f and ν to be correlated with s and x. The coefficients in the standard fraction female regressions do not generally have the same interpretation. Thirdly, we define the segregation contributions as differences in the mean values of the firm and job effects between men and women.14

Despite the differences in the modelling framework, the estimating wage equation in (10) and the associated decomposition are not much different from those in the previous studies. If occupational segregation is omitted, the standard fraction female decompositions can be viewed as special cases of our CRE decomposition. If we set δ1, δ2, θ1, and θ2 to zero, we obtain a specification similar to those in Datta Gupta and Rothstein (2001) and Bayard et al. (2003). If we impose further β = 0, the model is reduced to Groshen’s (1991) specification. Within our framework, the restriction δ1 = θ1 = 0 is equivalent to assuming that the firm and job effects are uncorrelated with x. This of course is a rather restrictive assumption, and it may lead to inconsistent estimates of η and β. The importance of this sort of restrictions is an empirical issue, and it depends on the data in hand. For example, both Datta Gupta and Rothstein (2001) and Bayard et al. (2003) find only a minor change in the female dummy coefficient when the fraction female variables were replaced with the full set of job dummies. In general, it does make a difference whether one conditions on the job held or only on the femaleness of the worker’s position. In our application, we find quantitatively significant discrepancies in the estimated coefficients, standard

14 Additional, less important, differences are: (1) we measure the fraction female in job as a deviation

from the firm mean, (2) we do not include the fraction female in occupation nor in industry in our model, and (3) we apply GLS, not OLS.

(36)

errors, and decomposition results between the CRE model and standard fraction female specification.

2.3 Data and descriptive statistics

2.3.1 The TT data and job classification

Our data come from the records of the Confederation of Finnish Industry and Employers (TT). TT is the central organisation of manufacturing employers and its member firms account for more than three-quarters of the value added of the Finnish manufacturing sector. Each year TT conducts three surveys covering almost all employees of its member firms. All surveys are directed to the employer, one asking information about white-collar workers and the other two about blue-collar workers. The focus of this paper is restricted to white-collar workers because of differences in the available records and compensation schemes between the two worker groups.15 I

References

Related documents

solution in the form of a polynomial. Recently, a great deal of effort has been done to apply this method for solving differ- ent types of differential equations [1,16] ,

The expected result will be the web-based GIS prototype using open specifications and open source software to manage the information by the web interface, and to present

In fact, NGS tech- nologies can be easily applied to a wide range of research fields; thus far, they have been successfully used to analyze target regions of the human genome ranging

● Starting Tuesday, October 13th, Blended Learning will allow students and staff the opportunity to spend 1-2 days per week in the school building to support work done at home..

The large intestine of the human consists of the cecum, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum. Ascending Colon Descending Colon

While the consultant focused primarily on the need for major structural changes in the State’s off-track betting operations - and its horse racing industry as a whole, it also

Correlations between online homework average and exam average, final exam score, final numerical grade, and final numerical grade less online homework contribution for

Whilst the role and relationships of social workers and health visitors is neither identical nor symmetrical, the work that they share in child protection generates